CMPT 454

1. Data Storage and Disk Access CMPT 454

2. Data Storage and Disk Access • Memory hierarchy • Hard disks • Architecture • Processing requests • Writing to disk • Hard disk reliability and efficiency • RAID • Solid State Drives • Buffer management • Data storage

3. Memory

4. DBMS and Memory Cache Main Memory DBMS Disk Virtual Memory File System Tertiary Storage

5. Memory Hierarchy • Primary memory: volatile • Main memory • Cache • Secondary memory: non-volatile • Solid State Drive (SSD) • Magnetic Disk (Hard Disk Drive, HDD) • Tertiary memory: non-volatile • CD/DVD • Tape - sequential access • Usually used as backup or for long-term storage cost speed

6. Main Memory vs Secondary Storage • Speed • Main memory is much faster than secondary memory • 10 – 100 nanoseconds to move data in main memory • 0.00001 to 0.0001 milliseconds • 10 milliseconds to read a block from an HDD • 0.1 milliseconds to read a block from a SSD • Cost • Main memory is around 100 times more expensive than secondary memory • SSDs are more expensive than HDDs

7. Main Memory vs Secondary Storage • System limitations • On a 32 bit system only 232 bytes can be directly referenced • Many databases are larger than that • Volatility • Data must be maintained between program executions which requires non-volatile memory • Nonvolatile storage retains its contents when the device is turned off, or if there is a power failure • Main memory is volatile, secondary storage is not

8. Hard Disk Drives

9. Magnetic Disks • Database data is usually stored on disks • A database will often be too large to be retained in main memory • When a query is processed data will need to be retrieved from storage • Data is stored on disk blocks • Also referred to as blocks, or in relation to the OS, pages • A contiguous sequence of bytes and • The unit in which data is written to, and read from • Block size is typically between 4 and 16 kilobytes

10. Magnetic Disk Structure • A hard disk consists of a number of platters • Platters can store data on either one or both of its surfaces so is referred to as • Single-sided or double sided • Surfaces are composed of concentric rings called tracks • The set of all tracks with the same diameter is called a cylinder • Sectors are arcs of a track • And are typically 4 kilobytes in size • Block size is set when the disk is initialized, usually a small multiple of the sector size (hence 4 to 16 kilobytes)

11. Diagram of a Disk surfaces platter 3* 2* track cylinder *statistics for Western Digital Caviar Black 1 TB hard drive

12. Disk Heads • Data is transferred to or from a surface by a disk head • There is one disk head for each surface • These disk heads are moved as a unit (called a disk head array) • Therefore all the heads are in identical positions with respect to their surfaces • To read or write a block a disk head must be positioned over it • Only one disk head can read or write at a time

13. Disk Anatomy the disk spins – around 7,200rpm disk head array track moves in and out platters

14. Disk Controller • Disk drives are controlled by a processor called a disk controller which • Controls the actuator that moves the head assembly • Selects sectors and determines when the disk has rotated to a sector • Transfers data between the disk and main memory • Some controllers buffer data from tracks in the expectation that the data will be required

15. Accessing Data in a Disk The disk constantly spins 7,200 rpm* The head pivots over the desired track The desired block is read as it passes underneath the head * Western Digital Caviar Black 1 TB hard drive (again)

16. Accessing A Block • The disk head is moved in or out to the track • This seek time is typically  10 milliseconds • WD Caviar Caviar Black 1TB: 8.9 ms • Wait until the block rotates under the disk head • This rotational delay is typically  4 milliseconds • WD Caviar Caviar Black 1TB : 4.2 ms • The data on the block is transferred to memory • This transfer time is the time it takes for the block to completely rotate past the disk head • Typically less than 1 millisecond

17. Transfer Time • The seek time and rotational delay depend on • Where the disk head is before the request, • Which track is being requested, and • How far the disk has to rotate • The transfer time depends on the request size • The transfer time (in ms) for one block equals • (60,000 / disk rpm) / blocks per track • The transfer time (in ms) for an entire track equals • (60,000 / disk rpm)

18. Main Memory versus Disk • Typical access time for a block on a hard disk • 15 milliseconds • Typical access time for a main memory frame • 60 nanoseconds • What’s the difference? • 1 millisecond = 1,000,000 nanoseconds • 60 ns = 0.000,060 ms • Accessing a hard drive is around 250,000 times slower than accessing main memory

19. Reducing Disk Access Time • Disk latency (access time) has three components • seek time + rotational delay + transfer time • The overall access time can be shortened by reducing, or even eliminating seek time and rotational delay • Related data should be stored in close proximity • Accessing two records in adjacent blocks on a track • Seek the desired track, rotate to first block, and transfer two blocks = 10 + 4 + 2*1 = 16ms • Accessing two records on different tracks • Seek the desired track, rotate to the block, and transfer the block, then repeat = (10 + 4 + 1)*2 = 30ms

20. Order of Closeness • What does it mean to say that related data should be stored close to each other? • The term close refers not to physical proximity but to how the access time is affected • In order of closeness: • Same block • Adjacent blocks on the same track • Same track • Same cylinder, but different surfaces • Adjacent cylinders • …

21. Which is Closer • Is 2, or 3 "closer" to 1? • 2 is in the adjacent track • And is clearly physically closer, but • The disk head must be moved to access it • 3 is in the same cylinder • The disk head does not have to be moved • Which is why 3 is closer 1 x x 2 x 3

22. Fulfilling Disk Requests • A fairalgorithm would take a first-come, first-serve approach • Insert requests in a queue and process them in the order in which they are received

23. Elevator Algorithm • The elevator algorithm usually performs better than FIFO • Requests are buffered and the disk head moves in one direction, processing requests • The arm then reverses direction

24. Requests – Discussion • The elevator algorithm gives much better performance than FIFO on average • And is a relatively fairalgorithm • The elevator algorithm is not optimal • The shortest-seek first algorithm is closer to optimal but can result in a high variance in response time • And may even result in starvation for distant requests • In some cases the elevator algorithm can perform worse than FIFO

25. Modifying a Record • To modify an existing record (on a disk) the following steps must be taken • Read the record • Modify the record in main memory • Write the modified record back to disk • It is important to remember that the smallest unit of transfer to / from a disk is a block • A single disk block usually contains many records

26. Read – Modify – Write Cycle Read one block into main memory …

27. Read – Modify – Write Cycle Read one block into main memory … … modify the desired record …

28. Read – Modify – Write Cycle Read one block into main memory … … modify the desired record … … and write it back.

29. Inserting Records • Consider creating a new record • The user enters the data for the record • Through some application interface • The record is created in main memory • And then written to disk • Does this process require a read-modify-write process? • YES! • Because, otherwise, the existing contents of the disk block will be overwritten

30. Disk Failures • Intermittent failure • Multiple attempts are required to read or write a sector • Media decay • A bit or a number of bits are permanently corrupted and it is impossible to read a sector • Write failure • A sector cannot be written to or retrieved • Often caused by a power failure during a write • Disk crash • The entire disk becomes unreadable

31. Checksums • An intermittent failure may result in incorrect data being read by the disk controller • Such incorrect data can be detected by a checksum • Each sector contains additional bits whose values are based on the data bits in the sector • A simple single-bit checksum is to maintain an even parity on the sector • If there is an odd number of 1s the parity is odd • If there is an even number of 1s the parity is even

32. Parity Bits • Assume that there are seven data bits and a single checksum bit • Data bits 0111011 – parity is odd • Checksum bit is set to 1 so that the overall parity is even • Using a single checksum bit allows errors of only one bit to be detected reliably • Several checksum bits can be maintained to reduce the chance of failing to notice an error • e.g. maintain 8 checksum bits, one for each bit position in the data bytes

33. Stable Storage • Checksums can detect errors but can't correct them • Stable storage can be implemented on a disk to allow errors to be corrected • Sectors are paired, with each pair representing a single sector • Pairs are usually referred to as Left and Right • Errors in a sector (L or R) are detected using checksums • Stable storage can cope with media failures and write failures

34. Stable Storage Policy • For writing, write the value of some sector X into XL • Check that the value is correct (using checksums) • If the value is not correct after a given number of attempts then assume that the sector has failed • A spare sector should be substituted for XL • Repeat the process for XR • For reading, XL and XR are read from in turn until a correct value is returned

35. RAID

36. Problems with Hard Disks • Hard disks act as bottlenecks for processing • DB data is stored on disks, and must be fetched into main memory to be processed, and • Disk access is considerably slower than main memory processing • There are also reliability issues with disks • Disks contain mechanical components that are more prone to failure than electronic components • One solution is to use multiple disks

37. Multiple Disks • Multiple disks • Each disk contains multiple platters • Disks can be read in parallel, and • Different disks can read from different cylinders • e.g. the first disk can access data from cylinder 6,000, while the second disk is accessing data from cylinder 11,000 • Single disk • Multiple platters • Disk heads are always over the same cylinder

38. Improving Efficiency • Using multiple disks to store data improves efficiency as the disks can be read in parallel • To satisfy a request the physical disks and disk blocks that the data resides on must be identified • The data may be on a single disk, or it may be split over multiple disks • The way in which data is distributed over the disks affects the cost of accessing it • In the same way that related data should be stored close to each other on a single disk

39. Data Striping • A disk array gives the user the abstraction of a single, large, disk • When an I/O request is issued the physical disk blocks to be retrieved have to be identified • How the data is distributed over the disks in the array affects how many disks are involved in an I/O request • Data is divided into partitions called striping units • The striping unit is usually either a block or a bit • Striping units are distributed over the disks using a round robin algorithm

40. disk 1 disk 2 disk 3 disk 4 Striping Notional File – the data is divided into striping units of a given size The striping units are distributed across a RAID system in a round robin fashion The size of the striping unit has an impact on the behaviour of the system

41. Striping Units – Block Striping Assume that a file is to be distributed across a four disk RAID system, using block striping, and that, Purely for the sake of illustration, the block size is just one byte! Notional File – the numbers represent a sequence of individual bits in the file Distribute these bits across a 4 disk RAID system using BLOCK striping: Disk 1 Disk 2 Disk 3 Disk 4 Block 1 Block 2 Block 3

42. Striping Units – Bit Striping Here is the same file to be distributed across a four disk RAID system, this time using bit striping, and again remember that Purely for the sake of illustration , the block size is just one byte! Notional File – the numbers represent a sequence of individual bits in the file Distribute these bits across a 4 disk RAID system using BIT striping: Disk 1 Disk 2 Disk 3 Disk 4 Block 3 Block 1 Block 2

43. Disk Array Performance • Assume that a disk array consists of D disks • Data is distributed across the disks using data striping • How does it perform compared to a single disk? • To answer this question we must specify the kinds of requests that will be made • Random read – reading multiple, unrelated records • Random write • Sequential read – reading a number of records (such as one file or table), stored on more than D blocks • Sequential write

44. The Basic Idea … • Use all D disks to improve efficiency, and distribute data using block striping • Random read performance • Very good – up to D different records can be read at once • Depending on which disks the records reside on • Random write performance – same as read performance • Sequential read performance • Very good – as related data are distributed over all D disks performance is D times faster than a single disk • Sequential write performance – same as read performance • But what about reliability …

45. Reliability • Hard disks contain mechanical components and are less reliable than other, purely electronic, components • Increasing the number of hard disks decreases reliability, reducing the mean-time-to-failure (MTTF) • The MTTF of a hard disk is  50,000 hours, or 5.7 years • In a disk array the overall MTTF decreases • Because the number of disks is greater • MTTF of a 100 disk array is 21 days – (50,000/100) / 24 • This assumes that failures occur independently and • The failure probability does not change over time • Reliability is improved by storing redundant data

46. Redundancy • Reliability of a disk array can be improved by storing redundant data • If a disk fails the redundant data can be used to reconstruct the data lost on the failed disk • The data can either be stored on a separate check disk or • Distributed uniformly over all the disks • Redundant data is typically stored using one of two methods • Mirroring, where each disk is duplicated • A parity scheme, where sufficient redundant data is maintained to recreate the data in any one disk • Other redundancy schemes provide greater reliability

47. Parity Scheme • For each bit on the data disks there is a parity bit on a check disk • If the sum of the data disks bits is even the parity bit is set to zero • If the sum of the bits is odd the parity bit is set to one • The data on any one failed disk can be recreated bit by bit 4 data disk system showing individual bit values 5th check disk containing parity data

48. Parity Scheme Read and Write • Reading • The parity scheme does not affect reading • Writing • A naïve approach would be to calculate the new value of the parity bit from all the data disks • A better approach is to compare the old and new values of the disk that is written to • And change the value of a parity bit if the corresponding bits have changed

49. Introducing RAID • A RAID system consists of several disks organized to increase performance and improve reliability • Performance is improved through data striping • Reliability is improved through redundancy • RAID stands for Redundant Arrays of Independent Disks • There are several RAID schemes or levels • The levels differ in terms of their • Read and write performance, • Reliability, and • Cost

50. RAID Level 0 • All D disks are used to improve efficiency, and data is distributed using block striping • No redundant information is kept • Read and write performance is very good • But, reliability is poor • Unless data is regularly backed up a RAID 0 system should only be used when the data is not important • A RAID 0 system is the cheapest of all RAID levels • As there are no disks used for storing redundant data